Calidaw

I wanted to explore music software where an agent could do more than offer advice or hand back a finished audio file. Could it work on the same material as the musician—notes, clips, instruments and effects—and leave behind something worth developing?

That makes control part of the creative problem. A useful idea should survive outside a chat. A change should have a clear destination. And when an experiment is wrong, getting back should be an ordinary Undo.

Calidaw grew from a native live looper into a larger studio: arrange clips, edit MIDI, build instrument sounds, route effects and mix a song. Desktop workspaces can split into different views or open in peer windows while sharing the same session.

The application has its own C++ audio engine and musical project model. The Mac interface uses native controls; composition tools sit alongside the timeline and instruments. Saving the project preserves the musical objects, rather than relying on the conversation that created them.

A person can turn a control in the interface. An agent can discover that control, read its current value and request a change. Both reach the same validated command path, with the same musical identities and Undo behaviour.

An agent first discovers what the running application supports. Changes name their target session and expected revision, so a delayed request cannot quietly edit a different song or overwrite a newer decision. Each applied operation returns a receipt; retrying the same operation does not create a second copy.

This is also an interaction-design choice. The visible application remains the place to inspect and continue the work. The agent does not own a parallel version of the song.

Phrase Lab is a shared composition workspace for harmony, motifs, runs, gestures and layered parts. A seed makes a generated idea reproducible; locks and explicit destinations help preserve what already works.

The visual editor and native agent access use the same deterministic engine and draft. A phrase can be previewed and planned before becoming MIDI clips and an optional named section in the song. Applying it is one musical Undo.

The recipe and the result serve different purposes. MIDI is the material a musician edits; the recipe retains the choices that produced it. Keeping those separate makes it possible to develop a phrase without pretending every manual edit is still generated.

A long list of parameters is technically complete and often musically unhelpful. I have been replacing that starting point with instrument-specific surfaces: a string model exposes contact and resonance; other instruments make their own sound-shaping relationships visible.

Racks bring instruments, effects and modulation into one saved structure. Their front, back and mixer are different views of the same graph. Turn the rack around and the connections are explicit, including note signals and one device modulating another.

The newer sound palette starts with an editable sound and a separate musical example. Directions such as warmth, bite and movement resolve to inspectable parameter changes. Audition and A/B comparison happen before applying a candidate to the song.

Copperlight began as a 24-bar study with six instruments, editable patterns and a five-section arrangement. Space & Motion is a separate mix that adds room, hall and echo returns, track treatment and section-aware automation.

The 53-second output below was rendered by Calidaw. The underlying notes, sounds, routing and automation remain editable. It is a development study, and a more useful demonstration than an empty editor or a list of implemented features.

The technical checks include restoring the complete mix with Undo and Redo, reopening the saved project independently, and rendering without clipped samples. Those checks matter; the musical result is something to hear and judge.

Using the application to compose exposed problems that isolated feature tests did not. A song could contain the right notes and still feel repetitive. An agent could report successful playback without having evaluated the result. A sound editor could expose every control while making the important ones hard to find.

That pushed the work toward motif development, deliberate arrangement, clearer sound controls and explicit listening comparisons. It also changed the way I evaluate progress: source tests, a running application, a saved song and a musical judgement answer different questions.

The next work continues that loop: improve the tools, make something with them, inspect what happened and let the result challenge the design.